(Note: An edited version of this column appeared in EETimes, May 14, 2007).
We live in a world of miniaturization and personal power. From desktop publishing and printing in the 1970s, we've gone to individual cell phones, to your own portable music library via MP3 players, and to full-capability desktop audio and video editing, to cite just a few. You can even buy a small-scale machine shop with CAD software which will route, drill, and turn parts for you with minimal intervention. If that won't cut it for you, so to speak, desktop stereolithography (rapid prototyping) machines will create high-quality parts for you, layer by layer, out of powered metal, various resins, or other materials, for both prototyping and small production runs.
But one important technology—semiconductor fabrication facilities–have not only resisted, but run counter to, the smaller, cheaper, do-it-yourself trend. They have increased in cost by several orders of magnitude over the past few decades. In this sense they are like their cousins, the particle physics labs, where ever-larger machines search for ever smaller particles.
Of course, to make these fabs financially viable, they have to run large volumes, reach for 100% utilization and yield, and minimize odd or special runs. Orders have to be carefully queued and scheduled to keep production moving, and the product flowing. The result is that specialty designs, test and investigative layouts, and student projects have nowhere to go, or have use special fabs that specialize in low volume, which is costly and time consuming. While programmable logic and final-layer metallization approaches help fill the gap, they are not suitable for many applications.
But what about taking a radically different approach to the problem of huge fabs? Suppose we set a challenge, perhaps similar to the Kremer prize and Ansari X-Prize in aviation, and get some student teams to build a practical desktop fab for low-volume, low-throughput production? By setting some reasonable constraints and boundaries, it might be doable. Once you redefine the objective and change the ground rules, then radical, new ideas for implementation may become apparent.
If the fab's objective is to make only a few parts, perhaps it should lock the wafer in place, and then robotically bring the process steps to the wafer in sequence. Maybe it doesn't use masks, but direct beam-writing. Limit wafer size to three inches. Have critical gases and fluids come pre-packaged, in small canisters, like ink for a desktop printer. Restrict the die size and number of package leads, as well as the packaging options. Support only a few processes, and don't go below 100 nm geometry (most designs don't need tighter geometry for performance; they need it primarily to achieve die per wafer density). Back-end processing could be done by an adjacent desktop unit.
This desktop fab would suit the needs of education, low-volume production runs, and some industry niches. Imagine being able to make a replacement part, just in time, for an obsolete IC that's failed (like the Star Trek replicator!). The challenge could be sparked with some prize money from the major foundries, capital equipment vendors, EDA houses, and even some of the IC companies, perhaps along with seed money to qualified teams.
We shouldn't continue to rely on multibillion-dollar fabs for prototype, low-volume, or specialty production runs. Instead, let's dramatically rescale the objectives, and perhaps we can get some ideas that make innovative use of today's small-scale, high-precision technologies. When you look at the problem from the outside in, “crazy” approaches may become very sensible.